def test_izhikevich_recurrent_cell_autapses(spiking_method):
cell = IzhikevichRecurrentCell(
2,
2,
spiking_method,
autapses=True,
recurrent_weights=torch.ones(2, 2) * 0.01,
dt=0.0001,
)
assert not torch.allclose(
torch.zeros(2),
(cell.recurrent_weights *
torch.eye(*cell.recurrent_weights.shape)).sum(0),
)
s1 = izhikevich.IzhikevichRecurrentState(z=torch.ones(1, 2),
v=torch.zeros(1, 2),
u=torch.zeros(1, 2))
_, s_full = cell(torch.zeros(1, 2), s1)
s2 = izhikevich.IzhikevichRecurrentState(
z=torch.tensor([[0, 1]], dtype=torch.float32),
v=torch.zeros(1, 2),
u=torch.zeros(1, 2),
)
_, s_part = cell(torch.zeros(1, 2), s2)
assert not s_full.v[0, 0] == s_part.v[0, 0]
def test_izhikevich_recurrent_cell_backward(spiking_method):
# Tests that gradient variables can be used in subsequent applications
cell = IzhikevichRecurrentCell(4, 4, spiking_method)
data = torch.randn(5, 4)
out, s = cell(data)
out, _ = cell(out, s)
loss = out.sum()
loss.backward()
def test_izhikevich_recurrent_cell(spiking_method):
cell = IzhikevichRecurrentCell(2, 4, spiking_method)
data = torch.randn(5, 2)
out, s = cell(data)
for x in s:
assert x.shape == (5, 4)
assert out.shape == (5, 4)
def test_izhikevich_recurrent_cell_no_autapses(spiking_method):
cell = IzhikevichRecurrentCell(2, 2, spiking_method, autapses=False)
assert (cell.recurrent_weights *
torch.eye(*cell.recurrent_weights.shape)).sum() == 0
s1 = izhikevich.IzhikevichRecurrentState(z=torch.ones(1, 2),
v=torch.zeros(1, 2),
u=torch.zeros(1, 2))
_, s_full = cell(torch.zeros(1, 2), s1)
s2 = izhikevich.IzhikevichRecurrentState(
z=torch.tensor([[0, 1]], dtype=torch.float32),
v=torch.zeros(1, 2),
u=torch.zeros(1, 2),
)
_, s_part = cell(torch.zeros(1, 2), s2)
assert s_full.v[0, 0] == s_part.v[0, 0]